Document interpreter LoRA

README.md

@@ -2,20 +2,36 @@
 license: apache-2.0
 base_model: mistralai/Ministral-8B-Instruct-2410
 tags:
-- mistral
 - lora
-- gradio
+- peft
+- mistral
 - interpretability
+- gradio
 - build-small-hackathon
 ---
 
 # Activation Brain Interpreter
 
-A LoRA fine-tuned Mistral-family interpreter for the Activation Brain Space.
+LoRA adapter for **Activation Brain**, trained to translate hidden-layer-derived telemetry from two Gemma-4-12B models into cautious plain-English comparison analysis.
 
-It is trained on hidden-layer-derived telemetry from two Gemma-4-12B models: prompt, dual responses, baseline-corrected emotion deltas, and model-native state meters. Its job is to generate plain-English comparison analysis without claiming that models literally feel human emotions.
+Base model: `mistralai/Ministral-8B-Instruct-2410`
 
-Space: https://huggingface.co/spaces/build-small-hackathon/activation-brain
-Artifacts/data: https://huggingface.co/datasets/build-small-hackathon/activation-brain-artifacts
+The adapter is used by the Activation Brain Space after both Gemma streams finish. It receives:
 
-Base model: `mistralai/Ministral-8B-Instruct-2410`
+- the user prompt
+- base Gemma response
+- OBLITERATED Gemma response
+- baseline-corrected emotion activation deltas
+- model-native state meters
+
+It returns compact JSON with:
+
+- `plain_english_read`
+- `what_changed`
+- `why_it_matters`
+- `best_takeaway`
+
+The model is trained to avoid claiming that language models literally feel emotions. It explains hidden-state-derived telemetry and response-tone differences.
+
+Space: https://huggingface.co/spaces/build-small-hackathon/activation-brain
+Artifacts: https://huggingface.co/datasets/build-small-hackathon/activation-brain-artifacts